Thank you very much James Gaskin... This is much more practical than just discussing it theoretically.. and you helps me a lot to understand this.. and for my thesis though.. Thanks again...
Hi James. Many thanks for the amazing video on MDA, particularly for the data visualisation part but I will have a correction about meeting the assumption of multivariate normality because Box's M does not test multivariate normality. It is used for testing the homogeneity of covariance matrices, which is also another assumption of discriminant analysis.
Nice video overall, by the way: at 8:00 this is actually not testing for multivariate normality, rather it is testing for equality of variance and covariance... but correct, you don’t want this to be significant...
Hi James! Your video helped me so much, i'm so grateful! I was wondering, do you remember where you found the information that 25% better than random was good enough?
oooh... It has been so long and I haven't used this since... My guess is that I found it in the Hair et al 2010 "multivariate data analysis" book. Good luck!
Thank you for your sharing. I would like to ask one question that how to get segmentation analysis for clothing market based on attitude and satisfaction? I would like to use multiple discriminant analysis.
Hi James, first a big thanks for lucidly explaining everything especially the plot. I am Professor Marketing and am doing research on Exploring Feminine traits from the shape of bottles. So 8 feminish bottles have to be marked on 7 adjectives on 7 point likert scale. Now I have to establish that these bottles have different traits. So I don't want to use discriminant for prediction but to know which of the independent variables are playing bigger role in discriminating between two bottles. Now for 2 levels of Nominal dependent variable we know that we can find it through standardized coefficients but I am stuck on how to know the same thing for 7 levels of dependent variable. Can you point from your analysis that from which part I should make the deduction?
I'm not sure if I understand, but you might try the two-step cluster analysis: th-cam.com/video/DpucueFsigA/w-d-xo.html And validation: th-cam.com/video/Odk0kLuUGvY/w-d-xo.html
Thank you so much! great material!! To predict observations (samples) as being from a specific class (group), the log of the likelihood ratios should be bigger than some threshold T, as described in some references, is there any threshold given in this method? Thanks in advance
James - great video. I have a question: in my segmentation i have used 4 factors + age+ personal income to determine 8 segments. How do Factors work with Discriminant Analysis?
if by factors you mean factor scores produced during an EFA, then they work the same as any continuous variable in MDA. Perhaps I have misunderstood the question.
Thanks James. I have created a segmentation using K means - the inputs are 4 factors + age + personal income. A second stage will involve running a DA to help us predict out of sample respondents into one of the segments created. This is where i became a little stuck - for any new respondent that answers the questions that made up the factor analysis, would i need to re-constitute the factor scores using the loadings before substituting into the DA Linear equations - so if you think of statements merging into factors and then factors forming the independents in the DA equations. Is that doable?
Maz Ali You should not have to run factor analysis again. It should be robust enough to handle new respondents. The only exception to this would be if you had a very small sample size to initially run the EFA. In this case, I would re-run the EFA to see if there are any differences.
Hi James. Thank you for uploading this video. I am currently doing cluster analysis followed by discriminant analysis. I am facing some issues and would be great if you could end my confusions. I get three clusters (assessed by hierarchical means and k - means clustering). When I do a discriminant analysis it gives two discriminant functions out of which the second discriminant function is non significant. What does this mean? Should that cluster be discarded or is it that due to small sample size of cluster it becomes non significant. What can be done to resolve this issue? Would be grateful if I get a response.
I have not used MDA since making this video five years ago, so I'm not sure I remember. Your suspicion of sample size in a cluster is a valid suspicion. Small sample size can lead to excessively influential error. As for what can be done, you can try to constrain it to fewer clusters (hence with larger sample sizes). Sorry I'm not more help on this one...
Hi James, Thanks for your video, it was very helpful! I'm hoping you could please provide me with the reference for Joe Hare's book, which you refer to in relation to your explanation of method choice (in this case Mahalanobis). Thank you in advance for your assistance! Georgia
Hi James. Thank you so much for this and the other great videos you post! I'm a fan! :) I have a question, what if my data is not normally distributed? and Box's M sig is .000? I've still tried to conduct the analysis you're showing and the final classification results are not bad at all (83.9% of original grouped cases correctly classified among 3 groups). So what to do?
You can list it as a limitation. If you have a ton of data, you can also just take smaller random samples of each group so that this statistic is not inflated.
Thank you very much James for sharing this. I've been looking for this function for very long time. The plot was not generated, though! why is that? I appreciate your help.
James Gaskin I did! When I checked the separate groups option, I got the plot. But when I check the combined group option, I don’t. Yet, I’m starting again from scratch tomorrow. Thank you for responding.
These functions are from the structure matrix. They are functionally similar to factor extracted during a factor analysis. They are something like proximities (or euclidean distances) to the centroid for a group of variables.
@ James Gaskin: Could you please post the references of your claim: discriminant function should at least predict 25 % better than random? Thanks in advance.
oooh. This was five years ago, and to be honest, I haven't used it since... So, I can't remember. It was probably from Hair et al 2010, but I don't have that book with me to check. I'd just have to start googling things like you. Best of luck to you.
I just saw in a comment from 3 years ago, that the reference is R.P. Burns, R. Burns, Business Research Methods and Statistics Using SPSS, SAGE Publications, Thousand Oaks, CA, USA, 2008
@@aliciachavez88 journals.sagepub.com/doi/full/10.1177/2515245919849378 gives good explanation of it. The 25% better than random seems to stem from Huberty I index which is the measure of improvement over chance rate. I=.25 (25%) is considered a medium -large effect depending on how many groups are present
Dear James. I have tried the method you thought in this video and it worked incredibly good. I have about 35000 samples and all of them in one graph is too much. Is there any way you recommend that I can reduce the number of points on the graph?
Wow! That's a lot of data! You could always just do a random sample. I would recommend doing a few random samples, and then comparing the MDA graphs to make sure they are consistent.
Sir, my question is on Multidimensional Scaling technique where in we have ranked data of certain banks on certain financial parameters for two different time period say 1999-2003 and 2006-2011, between which certain banks mergers happened. The purpose is to compare the performance of said banks across the merger period on the variables chosen using year wise ranked values on the financial parameters chosen. Can I use MDS technique for the above data which is a sort of small panel data, develop perception maps for the same
In fact, in continuation of my previous post, I submit that we had originally ranked data on certain proxy variables for each CAMEL inidcator bank-wise for both the periods, then we considered averages of the ranks for all proxy variables under each CAMEL parameter, to get year wise average rank values for each bank for each of the 5 such parameter for both the periods. Can we use MDS for comparing the banks across the merger using such averaged rank data computed for both the periods separately
Can I plot the discrimination function analysis if I have only two groups? From what I have tried, it's not possible, but I'm wondering if there's a way round it. Thanks in advance!
James Gaskin Many thanks James! Yes there’s no away around it, but if I use your trick of clearing out the group’s dots, I can present two groups only. It helps, but it doesn’t fully answer the question, but good for presentation purposes :)
MY Log Determinants table gives no value in the log determinant column and says with a note "a. Singular". The Box M test is saying , " NO test can be performed with fewer than two nonsingular group covariance matrices". What does those mean? Thanks in advance.
My guess is that one of the variables used in the analysis had zero variance. For example, if you used gender in the analysis, and all responded as male. Another possible reason might be extreme non-normality or lots of missing values.
Good video, James. Almost all discriminant analysis videos are too "complex" and you made it simple and on point. Thanks
Thank you very much James Gaskin... This is much more practical than just discussing it theoretically.. and you helps me a lot to understand this.. and for my thesis though.. Thanks again...
Hi James. Many thanks for the amazing video on MDA, particularly for the data visualisation part but I will have a correction about meeting the assumption of multivariate normality because Box's M does not test multivariate normality. It is used for testing the homogeneity of covariance matrices, which is also another assumption of discriminant analysis.
Nice video overall, by the way: at 8:00 this is actually not testing for multivariate normality, rather it is testing for equality of variance and covariance... but correct, you don’t want this to be significant...
Good to know! Thanks! I'm definitely not an expert on MDA. I learned it just for this video and a subsequent workshop I was teaching.
THIS MAKES SO MUCH SENSE THANK YOU!!
Thank you very much
Thank you sooo much for such a useful video, James!
Hi James! Your video helped me so much, i'm so grateful! I was wondering, do you remember where you found the information that 25% better than random was good enough?
oooh... It has been so long and I haven't used this since... My guess is that I found it in the Hair et al 2010 "multivariate data analysis" book. Good luck!
@@Gaskination Thank you so much!
Thank you for your sharing. I would like to ask one question that how to get segmentation analysis for clothing market based on attitude and satisfaction? I would like to use multiple discriminant analysis.
I'm not sure I understand your question. To do such a study, you would need to collect that data. Then, as you note, use MDA.
Hi James, first a big thanks for lucidly explaining everything especially the plot. I am Professor Marketing and am doing research on Exploring Feminine traits from the shape of bottles. So 8 feminish bottles have to be marked on 7 adjectives on 7 point likert scale. Now I have to establish that these bottles have different traits. So I don't want to use discriminant for prediction but to know which of the independent variables are playing bigger role in discriminating between two bottles. Now for 2 levels of Nominal dependent variable we know that we can find it through standardized coefficients but I am stuck on how to know the same thing for 7 levels of dependent variable. Can you point from your analysis that from which part I should make the deduction?
I'm not sure if I understand, but you might try the two-step cluster analysis: th-cam.com/video/DpucueFsigA/w-d-xo.html
And validation: th-cam.com/video/Odk0kLuUGvY/w-d-xo.html
Thank you so much! great material!! To predict observations (samples) as being from a specific class (group), the log of the likelihood ratios should be bigger than some threshold T, as described in some references, is there any threshold given in this method? Thanks in advance
Not that I'm aware of. Sorry about that. I'm not an expert on MDA though. I have only used it a few times.
@@Gaskination Thank you!!
nicely explained, thank you
James - great video. I have a question: in my segmentation i have used 4 factors + age+ personal income to determine 8 segments. How do Factors work with Discriminant Analysis?
if by factors you mean factor scores produced during an EFA, then they work the same as any continuous variable in MDA. Perhaps I have misunderstood the question.
Thanks James. I have created a segmentation using K means - the inputs are 4 factors + age + personal income. A second stage will involve running a DA to help us predict out of sample respondents into one of the segments created. This is where i became a little stuck - for any new respondent that answers the questions that made up the factor analysis, would i need to re-constitute the factor scores using the loadings before substituting into the DA Linear equations - so if you think of statements merging into factors and then factors forming the independents in the DA equations. Is that doable?
Maz Ali You should not have to run factor analysis again. It should be robust enough to handle new respondents. The only exception to this would be if you had a very small sample size to initially run the EFA. In this case, I would re-run the EFA to see if there are any differences.
Thanks James - you Rock!
Hi James. Thank you for uploading this video. I am currently doing cluster analysis followed by discriminant analysis. I am facing some issues and would be great if you could end my confusions. I get three clusters (assessed by hierarchical means and k - means clustering). When I do a discriminant analysis it gives two discriminant functions out of which the second discriminant function is non significant. What does this mean? Should that cluster be discarded or is it that due to small sample size of cluster it becomes non significant. What can be done to resolve this issue? Would be grateful if I get a response.
I have not used MDA since making this video five years ago, so I'm not sure I remember. Your suspicion of sample size in a cluster is a valid suspicion. Small sample size can lead to excessively influential error. As for what can be done, you can try to constrain it to fewer clusters (hence with larger sample sizes). Sorry I'm not more help on this one...
Hi James,
Thanks for your video, it was very helpful!
I'm hoping you could please provide me with the reference for Joe Hare's book, which you refer to in relation to your explanation of method choice (in this case Mahalanobis).
Thank you in advance for your assistance!
Georgia
Hair, J. F., Jr., Black, W. C., Babin, B. J., & Anderson, R. E. (2010). Multivariate data analysis (7th ed.). Upper Saddle River, NJ: Prentice Hall.
Thank you!
What reference(s) did you use that tells you "you want to be about 25% better than random"? 16:45
R.P. Burns, R. Burns, Business Research Methods and Statistics Using SPSS, SAGE Publications, Thousand Oaks, CA, USA, 2008.
Hi James. Thank you so much for this and the other great videos you post! I'm a fan! :)
I have a question, what if my data is not normally distributed? and Box's M sig is .000?
I've still tried to conduct the analysis you're showing and the final classification results are not bad at all (83.9% of original grouped cases correctly classified among 3 groups). So what to do?
You can list it as a limitation. If you have a ton of data, you can also just take smaller random samples of each group so that this statistic is not inflated.
Thank you very much James for sharing this. I've been looking for this function for very long time. The plot was not generated, though! why is that? I appreciate your help.
Make sure to check the box for "combined groups" plot in the classification options.
James Gaskin I did! When I checked the separate groups option, I got the plot. But when I check the combined group option, I don’t. Yet, I’m starting again from scratch tomorrow. Thank you for responding.
@@Gaskination It worked well today, many thanks again :)
Thanks JAMES, kindly i have a question that please explain assumptions of Multiple regression?
Multiple regression assumes univariate normality, linearity, and low-multicollinearity.
Hey James
Great video, thanks!
I have a question, what is the meaning of the axes in the graph ? what is it "function 1" and "function 2"?
These functions are from the structure matrix. They are functionally similar to factor extracted during a factor analysis. They are something like proximities (or euclidean distances) to the centroid for a group of variables.
My dear James, Can i get the spss database in order to practice the discriminat analysis. Regards
@ James Gaskin: Could you please post the references of your claim: discriminant function should at least predict 25 % better than random? Thanks in advance.
oooh. This was five years ago, and to be honest, I haven't used it since... So, I can't remember. It was probably from Hair et al 2010, but I don't have that book with me to check. I'd just have to start googling things like you. Best of luck to you.
Hi! Did you find the reference? I am also looking for it! :)
I just saw in a comment from 3 years ago, that the reference is R.P. Burns, R. Burns, Business Research Methods and Statistics Using SPSS, SAGE Publications, Thousand Oaks, CA, USA, 2008
@@aliciachavez88 journals.sagepub.com/doi/full/10.1177/2515245919849378 gives good explanation of it. The 25% better than random seems to stem from Huberty I index which is the measure of improvement over chance rate. I=.25 (25%) is considered a medium -large effect depending on how many groups are present
Hi Dear James.
Could you please post related ppt file like all other topics in SEM you did earlier.... regards
Here you go: drive.google.com/file/d/0B3T1TGdHG9aEb01CWXNQM0E5Qm8/view?usp=sharing
James Gaskin
Thanks for fast reply :)
hi james can we get the data file for practice
Dear James. I have tried the method you thought in this video and it worked incredibly good. I have about 35000 samples and all of them in one graph is too much. Is there any way you recommend that I can reduce the number of points on the graph?
Wow! That's a lot of data! You could always just do a random sample. I would recommend doing a few random samples, and then comparing the MDA graphs to make sure they are consistent.
Hi James. Could you please post the data set?
+TelsaBom Here is a link to the data: www.dropbox.com/s/58771yb3cgk5mu6/BurgersOriginal.sav?dl=0
Sir, my question is on Multidimensional Scaling technique where in we have ranked data of certain banks on certain financial parameters for two different time period say 1999-2003 and 2006-2011, between which certain banks mergers happened. The purpose is to compare the performance of said banks across the merger period on the variables chosen using year wise ranked values on the financial parameters chosen. Can I use MDS technique for the above data which is a sort of small panel data, develop perception maps for the same
In fact, in continuation of my previous post, I submit that we had originally ranked data on certain proxy variables for each CAMEL inidcator bank-wise for both the periods, then we considered averages of the ranks for all proxy variables under each CAMEL parameter, to get year wise average rank values for each bank for each of the 5 such parameter for both the periods. Can we use MDS for comparing the banks across the merger using such averaged rank data computed for both the periods separately
I'm not very good with MDS, so I'm not sure about this scenario. Sorry about that.
Can I plot the discrimination function analysis if I have only two groups? From what I have tried, it's not possible, but I'm wondering if there's a way round it. Thanks in advance!
Not that I'm aware of. Sorry about that.
James Gaskin
Many thanks James! Yes there’s no away around it, but if I use your trick of clearing out the group’s dots, I can present two groups only. It helps, but it doesn’t fully answer the question, but good for presentation purposes :)
MY Log Determinants table gives no value in the log determinant column and says with a note "a. Singular".
The Box M test is saying , " NO test can be performed with fewer than two nonsingular group covariance matrices".
What does those mean?
Thanks in advance.
My guess is that one of the variables used in the analysis had zero variance. For example, if you used gender in the analysis, and all responded as male. Another possible reason might be extreme non-normality or lots of missing values.
@@Gaskination the variables are non normal. I did a Shapiro wilk test and found the result to be significant.
Hi, dear James. Do me a favor please. can i get the spss database. Regards
Here you go: www.kolobkreations.com/BurgersOriginal.sav
I can get the spss database. regards
Here you go: www.kolobkreations.com/BurgersOriginal.sav